Evaluation of the Effect of Logarithmic Transformations and Objective Functions on the Performance of Neural Network Models in Estimation of Suspended Sediment Load (Case Study: Sarab Ghare So Watershed, Ghorichai River)

Accurate estimation of river suspended sediment load (SSL) has an important role in water resources, watershed management and related sciences. Due to the high fluctuations of SSL in different seasons of the year as well as its severely nonlinear and complexity nature, it is necessary to use appropriate methods that can simulate and estimate such phenomena.

In this study, data log transformation and multi-objective particle swarm optimization (MOPSO) algorithm were used for optimal training of neural network models. For this purpose, at first, by using unsupervised neural network (SOM), data of flow discharge and suspended sediment load of the studied hydrometric station (Statistical Period 1995-2016) were clustered. Then, by sampling the clusters, the data set needed to train and test the neural network models were prepared. After it, three scenarios were defined to evaluate the impact of applying logarithmic transformations and MOPSO optimization algorithm In the first scenario, the initial data (without logarithmic transformation) and the common base gradient algorithm in training neural network models (error propagation), in the second scenario, the error propagation algorithm and the logarithmic transforms, and in the third scenario, the logarithmic transforms and the MOPSO algorithm, was used to train neural network models.

Evaluation and comparison of the model validation results showed that applying a logarithmic transforms and MOPSO algorithm, by reducing RMSE error and bias percentage (PBIAS) from 49 ton/day and -21%, in the best model of the first scenario, to 30.3 ton/day and -6.3%, in the best model of scenario III, increased the efficiency of the models. Other results of the study are non-estimation of negative values for suspended sediment, which is one of the common errors in using neural network models in estimating suspended sediment.

The use of multiple objective functions makes it possible to sensitize the models to more accurately estimate the suspended sediment at low or high flows, thus improving the validation and skewness indices of the base data models.